WO2022052367A1 - Neural network optimization method for remote sensing image classification, and terminal and storage medium - Google Patents

Abstract

The present application relates to a neural network optimization method for remote sensing image classification, and a terminal and a storage medium. The method comprises: acquiring a remote sensing image data set; constructing an anti-noise network model, wherein the anti-noise network model comprises an image segmentation model and a loss selection model, and the image segmentation model is a U-Net network based on an SE module; and inputting the remote sensing image data set into the anti-noise network model for iterative training, performing, via the anti-noise network model, image segmentation by means of the U-Net network based on the SE module, so as to obtain an image classification result, using, via the loss selection model, a ksigma criterion to perform loss selection, and removing an error that exceeds a set deviation interval, so as to obtain an optimal network model parameter. By means of the embodiments of the present application, the feature extraction capability of a network model is improved, and the problem of a decrease in the classification precision of a neural network caused by noise of tags in a remote sensing image data set is solved.

Description

A neural network optimization method, terminal and storage medium for remote sensing image classification technical field

The application belongs to the technical field of remote sensing image processing, and in particular relates to a neural network optimization method, a terminal and a storage medium for remote sensing image classification.

Background technique

The classification problem of remote sensing images corresponds to the semantic segmentation problem in computer vision, which is to assign a classification category to each pixel in the image. At present, there is a noise problem in the data set labels in the remote sensing image classification process, mainly including more or less labeling of category pixels. Similar to the expansion or corrosion of the image, using a noisy data set to train the neural network will lead to the neural network. The classification performance is degraded and the obtained results are inaccurate.

There are two existing convolutional neural network algorithms for dealing with label noise. One is to model noise, build a noise processing model, update labels using network output results, and correct noisy labels during training. Another approach is to use a loss function that is robust to noise to improve the robustness of the neural network algorithm. The above algorithms can achieve good results in dealing with the problem of noisy labels in natural image classification, but they cannot be applied to the situation where the training labels are noisy.

With the great success of deep learning in the field of natural image processing, many researchers have applied the semantic segmentation methods in deep learning to remote sensing image classification and achieved good results. A crucial factor for deep learning to achieve superior results is to have an accurately labeled dataset as training learning. However, it is time-consuming and difficult to manually create an accurate and noise-free dataset in remote sensing images.

SUMMARY OF THE INVENTION

The present application provides a neural network optimization method, terminal, and storage medium for remote sensing image classification, aiming to solve one of the above-mentioned technical problems in the prior art at least to a certain extent.

In order to solve the above problems, the application provides the following technical solutions:

A neural network optimization method for remote sensing image classification, comprising:

Obtain remote sensing image datasets;

Build an anti-noise network model, the anti-noise network model includes an image segmentation model and a loss selection model, and the image segmentation model is a U-Net network based on the SE module;

Input the remote sensing image data set into the anti-noise network model for iterative training, and the anti-noise network model performs image segmentation through the U-Net network based on the SE module to obtain image classification results, and selects through the loss The model uses the ksigma criterion to select the loss, eliminates the error exceeding the set deviation interval, and obtains the optimal network model parameters.

The technical solution adopted in the embodiment of the present application further includes: the obtaining of the remote sensing image data set includes:

The remote sensing image data set is divided into training set, validation set and test set according to a set ratio, and the images of the training set, validation set and test set are cropped into images of a set size, and the training set images are Perform data cleaning and data enhancement.

The technical solutions adopted in the embodiments of the present application further include: performing image segmentation through the SE module-based U-Net network includes:

After the input feature map passes through a standard convolutional layer, two branches are generated. The first branch passes through two standard convolutional layers to obtain the first feature map; the second branch is the SE module, which includes a Globalpooling layer, two layers The FullyConnected layer and the sigmoid function layer firstly perform global average pooling on the input feature map through the Globalpooling layer to obtain the second feature map; and then activate the sigmoid function layer after passing through two Fully Connected layers to obtain the same feature as the second feature. The weight corresponding to the size of the image is multiplied by the first feature map generated by the first branch to obtain the image classification output result.

The technical solution adopted in the embodiment of the present application further includes: the loss selection using the ksigma criterion through the loss selection model includes:

If a set of test data roughly obeys a normal distribution and only contains random errors, the random errors are processed to obtain the standard deviation, and the deviation interval is determined according to the set probability, and the errors exceeding the deviation interval are determined as gross errors and eliminated. .

The technical solution adopted in the embodiment of the present application further includes: the inputting the remote sensing image dataset into the anti-noise network model for iterative training includes:

The training set is input into the anti-noise network model, the learning rate, the number of iterations, and the K value of the loss selection model are set, and the loss function for optimizing the network parameters is set, and the model training process is adjusted according to the loss curve.

The technical solution adopted in the embodiment of the present application further includes: the inputting the remote sensing image dataset into the anti-noise network model for iterative training further includes:

0%, 25% and 50% of the sample images are randomly selected from the training set, and 5*5, 7*7 and 9*9 convolution kernels are used to dilate and corrode the selected sample images to generate different types of and The noise-marked images of the level are trained according to the anti-noise network model according to the noise-marked images of different types and levels.

The technical solutions adopted in the embodiments of the present application further include: after obtaining the optimal network model parameters, the following further includes:

The test set image is input into the anti-noise network model, the classification result of the test set image is obtained, and the performance of the anti-noise network model is evaluated according to the classification result.

Another technical solution adopted by the embodiment of the present application is: a neural network optimization system, comprising:

Data acquisition module: used to acquire remote sensing image datasets;

Anti-noise network building module: used to construct an anti-noise network model, the anti-noise network model includes an image segmentation model and a loss selection model, and the image segmentation model is a U-Net network based on the SE module;

Model training module: used to input the remote sensing image data set into the anti-noise network model for iterative training, and the anti-noise network model performs image segmentation through the U-Net network based on the SE module to obtain an image classification result, And through the loss selection model, the ksigma criterion is used to select the loss, and the error exceeding the set deviation interval is eliminated to obtain the optimal network model parameters.

Another technical solution adopted by the embodiments of the present application is: a terminal, the terminal includes a processor and a memory coupled to the processor, wherein,

The memory stores program instructions for implementing the neural network optimization method for remote sensing image classification;

The processor is configured to execute the program instructions stored in the memory to control neural network optimization for remote sensing image classification.

Another technical solution adopted by the embodiments of the present application is: a storage medium storing program instructions executable by a processor, where the program instructions are used to execute the neural network optimization method for remote sensing image classification.

Compared with the prior art, the beneficial effects of the embodiments of the present application are: the neural network optimization method, system, terminal and storage medium for remote sensing image classification according to the embodiments of the present application improve the network model based on the semantic segmentation network U-Net , build an anti-noise network model, use the ksigma criterion for loss selection, add SE module to the anti-noise network model, improve the feature extraction ability of the network model, and solve the problem of neural network classification accuracy decline due to noise in labels in remote sensing image datasets. question.

Description of drawings

1 is a flowchart of a neural network optimization method for remote sensing image classification according to a first embodiment of the present application;

2 is an architecture diagram of an anti-noise network model according to an embodiment of the present application;

Fig. 3 is the existing U-Net network structure diagram;

Fig. 4 is the structure diagram of the SE module of the embodiment of the present application;

5 is a flowchart of a neural network optimization method for remote sensing image classification according to the second embodiment of the present application;

6 is a schematic structural diagram of a neural network optimization system for remote sensing image classification according to an embodiment of the application;

FIG. 7 is a schematic structural diagram of a terminal according to an embodiment of the present application;

FIG. 8 is a schematic structural diagram of a storage medium according to an embodiment of the present application.

detailed description

In order to make the purpose, technical solutions and advantages of the present application more clearly understood, the present application will be described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, but not to limit the present application.

Please refer to FIG. 1 , which is a flowchart of the neural network optimization method for remote sensing image classification according to the first embodiment of the present application. The neural network optimization method for remote sensing image classification according to the first embodiment of the present application includes the following steps:

S10: obtain a remote sensing image dataset;

Among them, the number of images and the size of the images in the remote sensing image dataset can be set according to the actual operation.

S11: Divide the remote sensing image dataset into a training set, a validation set and a test set according to a set ratio;

S12: Build an anti-noise network model based on SE module;

Among them, the anti-noise network model architecture is shown in Figure 2, which includes an image segmentation model and a loss selection model. The image segmentation model is a U-Net network based on the SE module. The network structure of the existing U-Net is shown in Figure 3, which includes two parts: a feature extraction part and an upsampling part. Among them, the feature extraction part is divided into five layers, and the image resolution is halved after each layer of pooling layer; correspondingly, the upsampling part is also divided into five layers, each of which has a standard volume containing two layers. The convolutional module of the stack. In this embodiment, the network model is improved on the basis of the existing U-Net, and the SE module (Squeeze-and-Excitation Networks) is added to the U-Net network structure to expand the perception of global information and improve the network's ability to deal with difficult problems. The learning ability of the sample.

Specifically, the improvement point of the network model in the embodiment of the present application is that the convolution module in the existing U-Net network structure is replaced by the SE module, which is used to improve the feature extraction capability of the network; the structure of the SE module is shown in FIG. 4 . Show. As shown in Figure 4, the image segmentation process of the image segmentation model is as follows: after inputting the feature map, it first passes through a standard convolution layer (Conv), and then generates two branches. The first branch passes through two standard convolution layers to obtain The first feature map of size C*3*3 (C is the feature map channel); the second branch is the SE module, including Globalpooling (global pooling layer), two layers of Fully Connected (full connection layer) and sigmoid function layer, First, global average pooling is performed on the input feature map through Globalpooling to obtain a second feature map with a size of C*1*1; then after two layers of Fully Connected (dimension reduction and dimension increase), it is activated by the sigmoid function layer to obtain the size is the weight of C*1*1, and the weight is multiplied by the first feature map generated by the first branch at the corresponding position to obtain the image classification output result.

In the process of network training, the loss obtained by samples with noisy labels will be larger than that obtained by samples with clean labels. Therefore, the loss selection model usually uses the ksigma algorithm to select the obtained losses and eliminate abnormal loss values. Thereby removing noise samples. However, when all high-loss samples are removed, the samples that are difficult to learn will also be removed. However, these samples that are difficult to learn play an important role in improving network performance. In view of this deficiency, in the embodiment of the present application, the loss selection model adopts the ksigma criterion to select the loss. It is assumed that a set of detection data roughly obeys the normal distribution and only contains random errors, and the random errors are processed to obtain the standard deviation, which is determined according to the set probability. A deviation interval, and the errors exceeding the deviation interval are determined as gross errors and eliminated.

S13: Input the training set into the anti-noise network model for iterative training to obtain optimal network model parameters;

S14: Input the test set into the trained anti-noise network model, obtain the classification result of the test set image, and evaluate the performance of the anti-noise network model according to the test result.

Based on the above, the neural network optimization method for remote sensing image classification according to the first embodiment of the present application uses the SE module to improve the semantic segmentation network U-Net, builds an anti-noise network model, improves the feature extraction capability of the network model, and uses ksigma Criterion for loss selection, to solve the problem of neural network classification accuracy decline due to noise in labels in remote sensing image datasets.

Please refer to FIG. 5 , which is a flowchart of the neural network optimization method for remote sensing image classification according to the second embodiment of the present application. The neural network optimization method for remote sensing image classification according to the second embodiment of the present application includes the following steps:

S20: Download Inria Aerial Image Labeling Dataset as a remote sensing image dataset;

Among them, this embodiment uses the Inria Aerial Image Labeling Dataset (which is a remote sensing image data set used for urban building detection) as the data set. The dataset includes a total of 180 remote sensing images with a size of 5000*5000 pixels. The annotation information of the dataset includes two types of buildings and non-buildings, which are mainly used for semantic segmentation.

S21: Construct training set, validation set and test set according to remote sensing image data set, at the same time crop the training set, validation set and test set images into images of a set size, and perform data cleaning and data enhancement operations on the training set images;

Among them, this embodiment only takes 135 images in the data set as the training set, 20 images as the validation set, and 25 images as the test set as examples, the three are independent of each other, and the images are randomly cropped into 256*256 images. , the specific image quantity and size can be set according to the actual operation. Data enhancement includes, but is not limited to, rotation, mirror symmetry, or/and adding Gaussian noise.

S22: Build an anti-noise network model based on the SE module;

S23: Input the training set into the anti-noise network model for training, and obtain the trained network model parameters;

Among them, the model training process is specifically: input the constructed training set into the anti-noise network model, set the hyperparameters such as the learning rate, the number of iterations, the K value of the loss selection model, and set the loss function used to optimize the network parameters. A good loss curve adjusts the training process, and finally gets the trained network model parameters.

Further, the embodiment of the present application randomly selects 0%, 25% and 50% of the sample images from the training set, and then uses 5*5, 7*7 and 9*9 convolution kernels to dilate and Corrosion is used to remove some noise samples to generate different types and levels of noise labeled images, and the anti-noise network model is trained according to the different types and levels of noise labeled images.

S24: Input the test set into the trained anti-noise network model, obtain the classification result of the test set image, and evaluate the performance of the anti-noise network model according to the classification result.

In order to verify the feasibility and effectiveness of the embodiments of the present application, the present application is tested through experiments below. The experiment uses pixel accuracy PA (Pixel Accuracy), average intersection ratio MIOU (Mean Intersection over Union), and Kappa coefficient as evaluation indicators, among which:

Among them, there are k+1 classes in total (from L0 to Lk, one of which is the background class), p _ij represents the number of pixels labeled as class i but predicted to be class j, p _ii indicates that the label is class i and the prediction is also a class The number of pixels in i, p _ji is the number of pixels labeled as class j but predicted to be class i, p _o is the sum of the number of correctly distributed samples for each class divided by the total number of samples, and p _e is the assumed number of each class The number of real samples is a1, a2 respectively, and the number of predicted samples of each class is b1, b2, and the total number of samples is n, then:

p _e = (a1*b1+a2*b2)/(n*n) (4)

By experimenting on the given dataset, the network is trained with different levels of noisy labels on the training set, tested on clean labels, and compared with existing U-Net networks. Table 1 below is the experimental results of the existing U-Net network and the anti-noise network model in the embodiment of the present application:

Table 1: The experimental results of the U-Net network and the anti-noise network model of this application

Dataset Noise Rate	noise type	method	PA	MIOU	Kappa
no noise	-	U-Net	0.919	0.714	0.763
no noise	-	anti-noise network	0.924	0.723	0.763
25% noise	Kernel5*5 corrosion	U-Net	0.923	0.718	0.743
25% noise	Kernel5*5 corrosion	anti-noise network	0.938	0.753	0.760
25% noise	Kernel7*7 Corrosion	U-Net	0.912	0.696	0.731
25% noise	Kernel7*7 Corrosion	anti-noise network	0.930	0.733	0.753
25% noise	Kernel9*9 Corrosion	U-Net	0.911	0.696	0.729
25% noise	Kernel9*9 Corrosion	anti-noise network	0.917	0.710	0.759
50% noise	Kernel5*5 corrosion	U-Net	0.909	0.689	0.722
50% noise	Kernel5*5 corrosion	anti-noise network	0.937	0.747	0.744
50% noise	Kernel7*7 Corrosion	U-Net	0.914	0.692	0.690
50% noise	Kernel7*7 Corrosion	anti-noise network	0.928	0.724	0.714
50% noise	Kernel9*9 Corrosion	U-Net	0.898	0.668	0.692
50% noise	Kernel9*9 Corrosion	anti-noise network	0.925	0.717	0.710
25% noise	Kernel5*5 expansion	U-Net	0.905	0.688	0.747
25% noise	Kernel5*5 expansion	anti-noise network	0.926	0.730	0.780
25% noise	Kernel7*7 expansion	U-Net	0.918	0.709	0.748
25% noise	Kernel7*7 expansion	anti-noise network	0.931	0.740	0.772
25% noise	Kernel9*9 inflation	U-Net	0.913	0.703	0.765
25% noise	Kernel9*9 inflation	anti-noise network	0.929	0.735	0.771
50% noise	Kernel5*5 corrosion	U-Net	0.906	0.686	0.744
50% noise	Kernel5*5 corrosion	anti-noise network	0.922	0.722	0.778
50% noise	Kernel7*7 Corrosion	U-Net	0.907	0.692	0.747
50% noise	Kernel7*7 Corrosion	anti-noise network	0.930	0.740	0.787
50% noise	Kernel9*9 Corrosion	U-Net	0.900	0.681	0.769
50% noise	Kernel9*9 Corrosion	anti-noise network	0.919	0.716	0.770

As can be seen from the above table, as the noise level increases in area and scale, the segmentation performance of the U-Net network decreases to varying degrees. On the other hand, the anti-noise network of the embodiment of the present application can maintain the same accuracy as no noise, even when the segmentation performance decreases slowly, even when the noise ratio is small. Therefore, the experimental results show that the embodiments of the present application can solve the problem that the classification accuracy of the neural network is reduced due to the existence of noise in the labels in the remote sensing image dataset.

Please refer to FIG. 6 , which is a schematic structural diagram of a neural network optimization system for remote sensing image classification according to an embodiment of the present application. The neural network optimization system for remote sensing image classification according to the embodiment of the present application includes:

Data acquisition module: used to acquire remote sensing image datasets;

Data segmentation module: It is used to divide the remote sensing image dataset into training set, validation set and test set according to the set ratio;

Anti-noise network building block: used to build an anti-noise network model;

Among them, the anti-noise network model includes an image segmentation model and a loss selection model. The image segmentation model is a U-Net network based on the SE module. The existing U-Net network structure includes two parts: the feature extraction part and the upsampling part. Among them, the feature extraction part is divided into five layers, and the image resolution is halved after each layer of pooling layer; correspondingly, the upsampling part is also divided into five layers, each of which has a standard volume containing two layers. The convolutional module of the stack. In this embodiment, the network model is improved on the basis of the existing U-Net, and the SE module (Squeeze-and-Excitation Networks) is added to the U-Net network structure to expand the perception of global information and improve the network's ability to deal with difficult problems. The learning ability of the sample.

Specifically, the improvement point of the network model in the embodiment of the present application is that the convolution module in the existing U-Net network structure is replaced by the SE module, which is used to improve the feature extraction capability of the network; the structure of the SE module is shown in FIG. 4 . The image segmentation process of the image segmentation model is as follows: after inputting the feature map, it first goes through a standard convolution layer (Conv), and then generates two branches. The first branch passes through two standard convolution layers, and the size is C* The first feature map of 3*3 (C is the feature map channel); the second branch is the SE module, including Globalpooling (global pooling layer), two layers of Fully Connected (full connection layer) and sigmoid function layer. The input feature map is subjected to global average pooling to obtain a second feature map of size C*1*1; then it is activated by the sigmoid function layer after two layers of Fully Connected (dimension reduction first and then dimension increase) to obtain a size of C*1 *1 weight, and multiply the weight with the first feature map generated by the first branch at the corresponding position to obtain the image classification output result.

In the process of network training, the loss obtained by samples with noisy labels will be larger than that obtained by samples with clean labels. Therefore, the loss selection model usually uses the ksigma algorithm to select the obtained losses and eliminate abnormal loss values. Thereby removing noise samples. However, when all the high-loss samples are removed, the samples that are difficult to learn will also be removed. However, these samples that are difficult to learn play an important role in improving network performance. In view of this deficiency, in the embodiment of the present application, the loss selection model adopts the ksigma criterion to select the loss. It is assumed that a set of detection data roughly obeys the normal distribution and only contains random errors, and the random errors are processed to obtain the standard deviation, which is determined according to the set probability. A deviation interval, and the errors exceeding the deviation interval are determined as gross errors and eliminated.

Model training module: used to input the training set into the anti-noise network model for training, and obtain the trained network model parameters;

Model evaluation module: It is used to input the test set into the trained anti-noise network model, obtain the classification results of the test set images, and evaluate the performance of the anti-noise network model according to the test results.

Please refer to FIG. 7 , which is a schematic structural diagram of a terminal according to an embodiment of the present application. The terminal 50 includes a processor 51 and a memory 52 coupled to the processor 51 .

The memory 52 stores program instructions for implementing the above-described neural network optimization method for remote sensing image classification.

The processor 51 is configured to execute program instructions stored in the memory 52 to control neural network optimization for remote sensing image classification.

The processor 51 may also be referred to as a CPU (Central Processing Unit, central processing unit). The processor 51 may be an integrated circuit chip with signal processing capability. The processor 51 may also be a general purpose processor, digital signal processor (DSP), application specific integrated circuit (ASIC), off-the-shelf programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware component . A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like.

Please refer to FIG. 8 , which is a schematic structural diagram of a storage medium according to an embodiment of the present application. The storage medium of this embodiment of the present application stores a program file 61 capable of implementing all the above methods, wherein the program file 61 may be stored in the above-mentioned storage medium in the form of a software product, and includes several instructions to enable a computer device (which may It is a personal computer, a server, or a network device, etc.) or a processor that executes all or part of the steps of the methods in the various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, Read-Only Memory (ROM, Read-Only Memory), Random Access Memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program codes , or terminal devices such as computers, servers, mobile phones, and tablets.

The above description of the disclosed embodiments enables any person skilled in the art to make or use the present application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined in this application may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (10)

1. A neural network optimization method for remote sensing image classification, comprising:

   Obtain remote sensing image datasets;

   Build an anti-noise network model, the anti-noise network model includes an image segmentation model and a loss selection model, and the image segmentation model is a U-Net network based on the SE module;

   Input the remote sensing image data set into the anti-noise network model for iterative training, and the anti-noise network model performs image segmentation through the U-Net network based on the SE module to obtain image classification results, and selects through the loss The model uses the ksigma criterion to select the loss, eliminates the error exceeding the set deviation interval, and obtains the optimal network model parameters.
2. The neural network optimization method for remote sensing image classification according to claim 1, wherein said acquiring a remote sensing image data set comprises:

   The remote sensing image data set is divided into training set, validation set and test set according to a set ratio, and the images of the training set, validation set and test set are cropped into images of a set size, and the training set images are Perform data cleaning and data enhancement.
3. The neural network optimization method for remote sensing image classification according to claim 1, wherein the performing image segmentation through the U-Net network based on the SE module comprises:

   After the input feature map passes through a standard convolutional layer, two branches are generated. The first branch passes through two standard convolutional layers to obtain the first feature map; the second branch is the SE module, which includes a Globalpooling layer, two layers The Fully Connected layer and the sigmoid function layer firstly perform global average pooling on the input feature map through the Globalpooling layer to obtain the second feature map; then after the two Fully Connected layers are activated by the sigmoid function layer, the second feature map is obtained. The weight corresponding to the size of the feature map is multiplied by the first feature map generated by the first branch to obtain the image classification output result.
4. The neural network optimization method for remote sensing image classification according to claim 3, wherein the loss selection using the ksigma criterion through the loss selection model comprises:

   If a set of test data roughly obeys a normal distribution and only contains random errors, the random errors are processed to obtain the standard deviation, and the deviation interval is determined according to the set probability, and the errors exceeding the deviation interval are determined as gross errors and eliminated. .
5. The neural network optimization method for remote sensing image classification according to claim 2, wherein the inputting the remote sensing image dataset into the anti-noise network model for iterative training comprises:

   The training set is input into the anti-noise network model, the learning rate, the number of iterations, and the K value of the loss selection model are set, and the loss function for optimizing network parameters is set, and the model training process is adjusted according to the loss curve.
6. The neural network optimization method for remote sensing image classification according to claim 5, wherein the inputting the remote sensing image dataset into the anti-noise network model for iterative training further comprises:

   0%, 25% and 50% of the sample images are randomly selected from the training set, and 5*5, 7*7 and 9*9 convolution kernels are used to dilate and corrode the selected sample images to generate different types of and The noise-marked images of the level are trained according to the noise-marked images of different types and levels, respectively.
7. The neural network optimization method for remote sensing image classification according to claim 2, wherein after obtaining the optimal network model parameters, the method further comprises:

   The test set image is input into the anti-noise network model, the classification result of the test set image is obtained, and the performance of the anti-noise network model is evaluated according to the classification result.
8. A neural network optimization system, characterized in that it includes:

   Data acquisition module: used to acquire remote sensing image datasets;

   Anti-noise network building module: used to build an anti-noise network model, the anti-noise network model includes an image segmentation model and a loss selection model, and the image segmentation model is a U-Net network based on the SE module;

   Model training module: used to input the remote sensing image data set into the anti-noise network model for iterative training, and the anti-noise network model performs image segmentation through the U-Net network based on the SE module to obtain an image classification result, And through the loss selection model, the ksigma criterion is used to select the loss, and the error exceeding the set deviation interval is eliminated to obtain the optimal network model parameters.
9. A terminal, characterized in that the terminal includes a processor and a memory coupled to the processor, wherein,

   The memory stores program instructions for realizing the neural network optimization method for remote sensing image classification according to any one of claims 1-7;

   The processor is configured to execute the program instructions stored in the memory to control neural network optimization for remote sensing image classification.
10. A storage medium, characterized in that it stores program instructions executable by a processor, and the program instructions are used to execute the neural network optimization method for remote sensing image classification according to any one of claims 1 to 7.

Images